Purchase this article with an account.
Jason Charng, Ignacio Viedma, David Alonso-Caneiro, David A Mackey, Fred Kuanfu Chen; A wavelength-agnostic, deep learning algorithm segmenting the hyperautofluorescent ring in retinitis pigmentosa. Invest. Ophthalmol. Vis. Sci. 2022;63(7):2073 – F0062.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Quantification of the hyperautofluorescent ring (HAR) is a key clinical trial outcome measure in retinitis pigmentosa (RP). Segmentation is currently done via traditional programming or manually. We applied a wavelength-agnostic deep learning (DL) algorithm to segment the HAR on fundus AF (FAF) images in RP eyes.
FAF images in RP patients were acquired using Heidelberg HRA2 in either blue or infrared AF (BAF /IRAF) mode, with a total of 1152 and 1099 images respectively.For each modality, a fully semantic algorithm (single model) based on a Unet++ architecture with an InceptionV2 encoder was developed and trained with an approximate 50/20/30% data split (training/validation/testing). The same network was then trained using both BAF and IRAF images (dual model) to produce a wavelength-agnostic platform. Model performance was assessed using the Dice similarity coefficient of the AF class, with 1 indicating perfect match to manual delineation. Paired t-test compared the difference in Dice score between the single and dual model.To test the wavelength-agnostic algorithm, we examined baseline FAF images from one eye (OD, if unavailable then OS) in patients with Usher syndrome. Manual and DL segmentation were performed in all images with HAR, with HAR horizontal extent and area extracted. Bland-Altman evaluated limits of agreement between manual and DL.
For the BAF dataset, comparable mean (standard deviation) Dice coefficient were found in the single 0.952 (0.042) and dual 0.953 (0.048) model (t-test, p=0.14). Single IRAF 0.957 (0.054) and dual 0.959 (0.053) model dice scores were also similar (p=0.24).In the Usher cohort (n=33), HAR was noted in 21 BAF and 22 IRAF images. In BAF, the mean difference (95% CI) between DL and manual delineation was 0.01 (-0.53 to 0.56) mm and -0.08 (-0.96 to 0.81) mm2 for horizontal ring extent and area, respectively. In 4 images, poor agreement between DL and manual segmentation was due to low image quality. In IRAF, the mean difference for ring extent was -0.29 (-2.73 to 2.15) mm and -0.89 (-8.59 to 6.81) mm2 for ring area. Poor agreement was due to low image quality (n=2) or indistinct HAR boundary (n=6).
The wavelength-agnostic DL algorithm was able to segment HAR in AF images, comparable to manual delineation, with manual correction required in a minority of images.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.
This PDF is available to Subscribers Only